Relieving label requirements through weakly supervised learning

Abstract

Precise annotation of data is a time-consuming and costly process, yet many models rely on having large amounts of labeled data for training. Many domains, including healthcare, have access to large amounts of unlabeled data, motivating groups to leverage this data for training their machine learning models. This thesis firstly investigates the correlation between the size of a medical dataset and the performance of models for different machine learning tasks, such as classification, detection, and segmentation. This thesis further explores the utility of training with additional unlabeled data by adopting various weakly supervised techniques. Finally, while comparing these techniques to supervised learning baselines, we highlight the trade-offs between using smaller labeled data sizes and model performances.